Where Should Learning Live? Regret Analysis for Text, Parameter, and Joint Adaptation
On Spider, text-only selection made fewer mistakes within its available choices yet produced more failures because stronger configurations were excluded.
The study compares selection among prebuilt prompts and parameter adapters against the best fixed pair used from the start. It separates excess loss into selection regret, restrictions on the available candidates, and the effects of deployment history. Under bounded history effects and full feedback on each candidate’s own-history loss, limiting switches yields sublinear expected selection regret. Controlled experiments show that fewer switches can help or hurt, depending on whether inherited histories are harmful or helpful.
ArXiv · AI/CL/LG's note
The study compares selection among prebuilt prompts and parameter adapters against the best fixed pair used from the start. It separates excess loss into selection regret, restrictions on the available candidates, and the effects of deployment history. Under bounded history effects and full feedback on each candidate’s own-history loss, limiting switches yields sublinear expected selection regret. Controlled experiments show that fewer switches can help or hurt, depending on whether inherited histories are harmful or helpful.
ArXiv · AI/CL/LG's note
score 4